Abstract:
The construction industry, essential for economic growth, is notably hazardous, with
falls from height (FFH) being the primary cause of injuries and fatalities. To enhance
construction safety ANN, which is a Machine Learning (ML) application, can be
employed. Typically, there are imbalances in accident data, for instance, more data is
available on hospitalization and less data available on fatalities. Such imbalances can
result in inaccurate findings from ML models. Past studies have addressed class
imbalance in accident data but lacked focus on its impact on ML algorithms. This study
considers fall-related accidents in construction by using recent data from OSHA to
develop a predictive model for injury severity. After preprocessing of data, critical
variables are identified using association analysis and an Artificial Neural Network
(ANN) is employed to capture the complex, nonlinear relationships between variables.
The study explores the impact of addressing class imbalance on the performance of
different architecture of ANN model through three resampling techniques: Random
under sampling (RUS), Random over sampling (ROS), and Synthetic Minority
oversampling technique (SMOTE). Resampled data is then compared with raw data.
Results indicate that worker age, occupation, fall distance, and working surface height
significantly impact injury severity, while environmental factors and nature of the task
do not significantly associate with accident outcomes. Simpler model architectures and
RUS are more effective and provide the best balance between precision and recall in
predicting injury severity. These findings are critical for the development of safetyrelated predictive models in future. This study will help relevant stakeholders such as
safety managers to take proactive steps by focusing on the critical variables identified
in analysis and to help them effectively manage fall accidents.
